Optical fan-out and broadcast interconnect

Abstract

Methods and apparatus are described for an optical fan-out and broadcast interconnect. A method includes operating an optical fan-out and broadcast interconnect including: fanning-out an optical signal from an optical signal emitter, of one of a plurality of nodes, with a diverging element of one of a plurality of optics; and broadcasting the optical signal to one of a plurality of receivers of all of the plurality of nodes with a light collecting and focusing element of all of the plurality of optics, wherein the plurality of optics are positioned to define an optics array, the plurality of receivers are positioned to define a receiver array that corresponds to the optics array and the plurality of nodes are positioned to define a node array that substantially corresponds to the receiver array and the optics array.

Description

[0001] This application claims a benefit of priority under 35 U.S.C. 119(e) from both copending provisional patent application U.S. Ser. No. 60 / 423,939, filed Nov. 5, 2002 and copending provisional patent application U.S. Ser. No. 60 / 432,141, filed Dec. 10, 2002, the entire contents of both of which are hereby expressly incorporated herein by reference for all purposes.[0002] 1. Field of the Invention[0003] The invention relates generally to the field of optical interconnects for computer systems and / or their subsystems as well as networks and / or their subsystems. More particularly, the invention relates to a free-space optical interconnect that includes a fan-out and broadcast signal link.[0004] 2. Discussion of the Related Art[0005] The concept of parallel-distributed processing (PDP), which is the theory and practice of massively parallel processing machines, predates the first supercomputers of the 1960s. In practice, high-performance parallel-distributed processing machines are...

AI Technical Summary

Problems solved by technology

In practice, high-performance parallel-distributed processing machines are difficult to achieve for several interrelated reasons.

The impact on software is roughly proportional to n.sup.2 due to the increased complexity of parallel-distributed processing algorithms.

The overall cost per node increases more rapidly than the number of nodes when all these factors are considered.

The total number of transistors is expected to be around 5 million, making for a large, expensive, and relatively power-hungry node.

Moreover, the main, unsolved problem facing today's supercomputers is how to achieve the economies of scale found elsewhere in the industrial world.

Part of the reason for this lack of progress in supercomputer scaling is that the interconnect problem has not yet found a satisfactory solution.

Adopting present solutions leads to a reliance on slow and bulky, off-chip hardware to carry the message traffic between processors.

A related problem is that communication delays increase as the number of nodes increases, meaning that the law of diminishing returns soon sets in.

However, using faster and more powerful nodes increases both the cost per node and the overall power consumption.

These programs run poorly on today's massively parallel machines.

Massively parallel high performance computers using fat tree and crossbar interconnect suffer from a mismatch with the software requirement for non-blocking broadcast of short messages.

Such broadcast uses excessive bandwidth in fat-tree interconnects which results in poor system performance.

Additionally, these all-to-all messages are typically short, being a few bytes in length.

Present systems can broadcast information, but only by simulating the broadcast function; thus their capability for implementing the all-to-all function is inefficient.

Of course, nothing prevents one from using the ultra-performance processors as nodes in the proposed systems; both cost and capability would rise significantly.

While theoretically alleviating the communications bottle-neck problem and helping to overcome data-dependency issues, the cure is literally worse than the disease since the nodes now spend more time managing the system's tasks in software than is gained by decomposing complex programs into tasks in the first place.

As data rates increase and data processors become faster, electrical communication between data-processing nodes becomes more power intensive and expensive.

As the number of processing nodes communicating within a system increases, electrical communication become slower due to increased distance and capacitance as well as more cumbersome due to the geometric increase in the number of wires, the volume of the crossbar, as well as its mass and power consumption.

Close proximity of communication channels produces crosstalk, which is perceived as noise on adjacent channels.

Neither of these problems occur in a light-based interconnect.

The main problem with today's optical solutions is conceptual: they are trying to solve a more complicated problem than necessary.

For an optical system serving hundreds of thousands of nodes, the mechanical alignment is an insurmountable nightmare.

The main problems with FSOI to overcome are alignment, where each laser must hit a specific receiver, and mechanical robustness.

Most FSOI methods lack direct broadcast capability due to the one-emitter, one-receiver assumption.

However, since the receivers are typically small devices, perhaps a tenth of a millimeter in diameter, it is difficult to achieve and maintain optical alignment of the narrow laser beam onto one or more receivers across all but the smallest distances.

This approach also suffers from sensitivity to alignment which is augmented by temperature sensitivity of the hologram material that affects the size of the fan-out pattern.

Images